From 0fd6c1f015f6cccf3b527f7dbd8386a434728126 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Thu, 14 Mar 2024 10:12:29 +0200 Subject: [PATCH 01/15] embedding : print cosine similarity (#899) --- common/common.cpp | 13 +++++++++++++ common/common.h | 1 + examples/embedding/embedding.cpp | 21 ++++++++++++++++----- examples/gritlm/gritlm.cpp | 26 ++++++-------------------- 4 files changed, 36 insertions(+), 25 deletions(-) diff --git a/common/common.cpp b/common/common.cpp index 73b1b61ba..58fbd05aa 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -1877,3 +1877,16 @@ void llama_embd_normalize(const float * inp, float * out, int n) { } } +float llama_embd_similarity_cos(const float * embd1, const float * embd2, int n){ + double sum = 0.0; + double sum1 = 0.0; + double sum2 = 0.0; + + for (int i = 0; i < n; i++) { + sum += embd1[i] * embd2[i]; + sum1 += embd1[i] * embd1[i]; + sum2 += embd2[i] * embd2[i]; + } + + return sum / (sqrt(sum1) * sqrt(sum2)); +} diff --git a/common/common.h b/common/common.h index 0f178b9eb..d250eef8b 100644 --- a/common/common.h +++ b/common/common.h @@ -268,3 +268,4 @@ void dump_kv_cache_view_seqs(const llama_kv_cache_view & view, int row_size = 40 void llama_embd_normalize(const float * inp, float * out, int n); +float llama_embd_similarity_cos(const float * embd1, const float * embd2, int n); diff --git a/examples/embedding/embedding.cpp b/examples/embedding/embedding.cpp index 49302a199..f390c4061 100644 --- a/examples/embedding/embedding.cpp +++ b/examples/embedding/embedding.cpp @@ -168,14 +168,25 @@ int main(int argc, char ** argv) { batch_decode(ctx, batch, out, s, n_embd); // print first 3 embeddings + fprintf(stdout, "\n"); for (int j = 0; j < std::min(3, n_prompts); j++) { - fprintf(stderr, "embedding %d: ", j); - for (int i = 0; i < n_embd; i++) { - fprintf(stderr, "%f ", emb[j * n_embd + i]); + fprintf(stdout, "embedding %d: ", j); + for (int i = 0; i < std::min(16, n_embd); i++) { + fprintf(stdout, "%f ", emb[j * n_embd + i]); } - fprintf(stderr, "\n\n"); + fprintf(stdout, "\n"); + } + + // print cosine similarity matrix + fprintf(stdout, "\n"); + printf("cosine similarity matrix:\n\n"); + for (int i = 0; i < n_prompts; i++) { + for (int j = 0; j < n_prompts; j++) { + float sim = llama_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd); + fprintf(stdout, "%6.2f ", sim); + } + fprintf(stdout, "\n"); } - fprintf(stderr, "\n"); // clean up llama_print_timings(ctx); diff --git a/examples/gritlm/gritlm.cpp b/examples/gritlm/gritlm.cpp index 3d4b085d6..52fd719b3 100644 --- a/examples/gritlm/gritlm.cpp +++ b/examples/gritlm/gritlm.cpp @@ -6,22 +6,6 @@ // #define GRIT_DEBUG -static float dot_product(const std::vector & v1, const std::vector & v2) { - float dot = 0.0f; - for (uint64_t i = 0; i < v1.size(); ++i) { - dot += v1[i] * v2[i]; - } - return dot; -} - -static float norm(const std::vector & v) { - return std::sqrt(dot_product(v, v)); -} - -static float cosine_similarity(const std::vector & v1, const std::vector & v2) { - return dot_product(v1, v2) / (norm(v1) * norm(v2)); -} - static std::vector> encode(llama_context * ctx, const std::vector & sentences, const std::string & instruction) { std::vector> result; @@ -203,10 +187,12 @@ int main(int argc, char * argv[]) { const std::vector> d_rep = encode(ctx, documents, gritlm_instruction("")); const std::vector> q_rep = encode(ctx, queries, gritlm_instruction(instruction)); - const float cosine_sim_q0_d0 = cosine_similarity(q_rep[0], d_rep[0]); - const float cosine_sim_q0_d1 = cosine_similarity(q_rep[0], d_rep[1]); - const float cosine_sim_q1_d0 = cosine_similarity(q_rep[1], d_rep[0]); - const float cosine_sim_q1_d1 = cosine_similarity(q_rep[1], d_rep[1]); + const int n_embd = llama_n_embd(mdl); + + const float cosine_sim_q0_d0 = llama_embd_similarity_cos(q_rep[0].data(), d_rep[0].data(), n_embd); + const float cosine_sim_q0_d1 = llama_embd_similarity_cos(q_rep[0].data(), d_rep[1].data(), n_embd); + const float cosine_sim_q1_d0 = llama_embd_similarity_cos(q_rep[1].data(), d_rep[0].data(), n_embd); + const float cosine_sim_q1_d1 = llama_embd_similarity_cos(q_rep[1].data(), d_rep[1].data(), n_embd); std::printf("Cosine similarity between \"%.50s\" and \"%.50s\" is: %.3f\n", queries[0].c_str(), documents[0].c_str(), cosine_sim_q0_d0); std::printf("Cosine similarity between \"%.50s\" and \"%.50s\" is: %.3f\n", queries[0].c_str(), documents[1].c_str(), cosine_sim_q0_d1); From 381da2d9f0940d7009e3e918bed36338c8ff2fbb Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Thu, 14 Mar 2024 11:55:23 +0200 Subject: [PATCH 02/15] metal : build metallib + fix embed path (#6015) * metal : build metallib + fix embed path ggml-ci * metal : fix embed build + update library load logic ggml-ci * metal : fix embeded library build ggml-ci * ci : fix iOS builds to use embedded library --- .github/workflows/build.yml | 2 ++ .gitignore | 2 ++ CMakeLists.txt | 70 ++++++++++++++++++++----------------- Makefile | 15 ++++---- ggml-metal.m | 49 +++++++++++++++++--------- ggml-metal.metal | 3 -- 6 files changed, 83 insertions(+), 58 deletions(-) diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml index d39cd6bc3..0da01d5ba 100644 --- a/.github/workflows/build.yml +++ b/.github/workflows/build.yml @@ -333,6 +333,7 @@ jobs: mkdir build cd build cmake -G Xcode .. \ + -DLLAMA_METAL_EMBED_LIBRARY=ON \ -DLLAMA_BUILD_EXAMPLES=OFF \ -DLLAMA_BUILD_TESTS=OFF \ -DLLAMA_BUILD_SERVER=OFF \ @@ -361,6 +362,7 @@ jobs: mkdir build cd build cmake -G Xcode .. \ + -DLLAMA_METAL_EMBED_LIBRARY=ON \ -DLLAMA_BUILD_EXAMPLES=OFF \ -DLLAMA_BUILD_TESTS=OFF \ -DLLAMA_BUILD_SERVER=OFF \ diff --git a/.gitignore b/.gitignore index d28f4d1b8..1ad8d929b 100644 --- a/.gitignore +++ b/.gitignore @@ -25,6 +25,8 @@ .vscode/ .idea/ +ggml-metal-embed.metal + lcov-report/ gcovr-report/ diff --git a/CMakeLists.txt b/CMakeLists.txt index a8abf4088..3ac2804a6 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -200,9 +200,6 @@ if (LLAMA_METAL) add_compile_definitions(GGML_METAL_NDEBUG) endif() - # get full path to the file - #add_compile_definitions(GGML_METAL_DIR_KERNELS="${CMAKE_CURRENT_SOURCE_DIR}/") - # copy ggml-common.h and ggml-metal.metal to bin directory configure_file(ggml-common.h ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h COPYONLY) configure_file(ggml-metal.metal ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal COPYONLY) @@ -211,53 +208,62 @@ if (LLAMA_METAL) enable_language(ASM) add_compile_definitions(GGML_METAL_EMBED_LIBRARY) + set(METALLIB_COMMON "${CMAKE_CURRENT_SOURCE_DIR}/ggml-common.h") set(METALLIB_SOURCE "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal.metal") + file(MAKE_DIRECTORY "${CMAKE_BINARY_DIR}/autogenerated") - set(EMBED_METALLIB_ASSEMBLY "${CMAKE_BINARY_DIR}/autogenerated/ggml-embed-metallib.s") + + # merge ggml-common.h and ggml-metal.metal into a single file + set(METALLIB_EMBED_ASM "${CMAKE_BINARY_DIR}/autogenerated/ggml-metal-embed.s") + set(METALLIB_SOURCE_EMBED "${CMAKE_BINARY_DIR}/autogenerated/ggml-metal-embed.metal") add_custom_command( - OUTPUT ${EMBED_METALLIB_ASSEMBLY} - COMMAND echo ".section __DATA,__ggml_metallib" > ${EMBED_METALLIB_ASSEMBLY} - COMMAND echo ".globl _ggml_metallib_start" >> ${EMBED_METALLIB_ASSEMBLY} - COMMAND echo "_ggml_metallib_start:" >> ${EMBED_METALLIB_ASSEMBLY} - COMMAND echo ".incbin \\\"${METALLIB_SOURCE}\\\"" >> ${EMBED_METALLIB_ASSEMBLY} - COMMAND echo ".globl _ggml_metallib_end" >> ${EMBED_METALLIB_ASSEMBLY} - COMMAND echo "_ggml_metallib_end:" >> ${EMBED_METALLIB_ASSEMBLY} - DEPENDS ${METALLIB_SOURCE} + OUTPUT ${METALLIB_EMBED_ASM} + COMMAND echo "Embedding Metal library" + COMMAND sed -e '/\#include \"ggml-common.h\"/r ${METALLIB_COMMON}' -e '/\#include \"ggml-common.h\"/d' < ${METALLIB_SOURCE} > ${METALLIB_SOURCE_EMBED} + COMMAND echo ".section __DATA,__ggml_metallib" > ${METALLIB_EMBED_ASM} + COMMAND echo ".globl _ggml_metallib_start" >> ${METALLIB_EMBED_ASM} + COMMAND echo "_ggml_metallib_start:" >> ${METALLIB_EMBED_ASM} + COMMAND echo ".incbin \\\"${METALLIB_SOURCE_EMBED}\\\"" >> ${METALLIB_EMBED_ASM} + COMMAND echo ".globl _ggml_metallib_end" >> ${METALLIB_EMBED_ASM} + COMMAND echo "_ggml_metallib_end:" >> ${METALLIB_EMBED_ASM} + DEPENDS ggml-metal.metal ggml-common.h COMMENT "Generate assembly for embedded Metal library" ) - set(GGML_SOURCES_METAL ${GGML_SOURCES_METAL} ${EMBED_METALLIB_ASSEMBLY}) - endif() - - if (LLAMA_METAL_SHADER_DEBUG) - # custom command to do the following: - # xcrun -sdk macosx metal -fno-fast-math -c ggml-metal.metal -o ggml-metal.air - # xcrun -sdk macosx metallib ggml-metal.air -o default.metallib - # - # note: this is the only way I found to disable fast-math in Metal. it's ugly, but at least it works - # disabling fast math is needed in order to pass tests/test-backend-ops - # note: adding -fno-inline fixes the tests when using MTL_SHADER_VALIDATION=1 - # note: unfortunately, we have to call it default.metallib instead of ggml.metallib - # ref: https://github.com/ggerganov/whisper.cpp/issues/1720 - set(XC_FLAGS -fno-fast-math -fno-inline -g) - if (LLAMA_QKK_64) - set(XC_FLAGS ${XC_FLAGS} -DQK_K=64) + set(GGML_SOURCES_METAL ${GGML_SOURCES_METAL} ${METALLIB_EMBED_ASM}) + else() + if (LLAMA_METAL_SHADER_DEBUG) + # custom command to do the following: + # xcrun -sdk macosx metal -fno-fast-math -c ggml-metal.metal -o ggml-metal.air + # xcrun -sdk macosx metallib ggml-metal.air -o default.metallib + # + # note: this is the only way I found to disable fast-math in Metal. it's ugly, but at least it works + # disabling fast math is needed in order to pass tests/test-backend-ops + # note: adding -fno-inline fixes the tests when using MTL_SHADER_VALIDATION=1 + # note: unfortunately, we have to call it default.metallib instead of ggml.metallib + # ref: https://github.com/ggerganov/whisper.cpp/issues/1720 + set(XC_FLAGS -fno-fast-math -fno-inline -g) + else() + set(XC_FLAGS -O3) endif() add_custom_command( OUTPUT ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib COMMAND xcrun -sdk macosx metal ${XC_FLAGS} -c ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.air COMMAND xcrun -sdk macosx metallib ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.air -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib - DEPENDS ggml-metal.metal + COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.air + COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h + COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal + DEPENDS ggml-metal.metal ggml-common.h COMMENT "Compiling Metal kernels" - ) + ) add_custom_target( ggml-metal ALL DEPENDS ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib - ) - endif() + ) + endif() # LLAMA_METAL_EMBED_LIBRARY set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} ${FOUNDATION_LIBRARY} diff --git a/Makefile b/Makefile index db9968efb..cb597b209 100644 --- a/Makefile +++ b/Makefile @@ -557,15 +557,16 @@ ggml-metal.o: ggml-metal.m ggml-metal.h $(CC) $(CFLAGS) -c $< -o $@ ifdef LLAMA_METAL_EMBED_LIBRARY -ggml-metal-embed.o: ggml-metal.metal +ggml-metal-embed.o: ggml-metal.metal ggml-common.h @echo "Embedding Metal library" + @sed -e '/#include "ggml-common.h"/r ggml-common.h' -e '/#include "ggml-common.h"/d' < ggml-metal.metal > ggml-metal-embed.metal $(eval TEMP_ASSEMBLY=$(shell mktemp)) - @echo ".section __DATA, __ggml_metallib" > $(TEMP_ASSEMBLY) - @echo ".globl _ggml_metallib_start" >> $(TEMP_ASSEMBLY) - @echo "_ggml_metallib_start:" >> $(TEMP_ASSEMBLY) - @echo ".incbin \"$<\"" >> $(TEMP_ASSEMBLY) - @echo ".globl _ggml_metallib_end" >> $(TEMP_ASSEMBLY) - @echo "_ggml_metallib_end:" >> $(TEMP_ASSEMBLY) + @echo ".section __DATA, __ggml_metallib" > $(TEMP_ASSEMBLY) + @echo ".globl _ggml_metallib_start" >> $(TEMP_ASSEMBLY) + @echo "_ggml_metallib_start:" >> $(TEMP_ASSEMBLY) + @echo ".incbin \"ggml-metal-embed.metal\"" >> $(TEMP_ASSEMBLY) + @echo ".globl _ggml_metallib_end" >> $(TEMP_ASSEMBLY) + @echo "_ggml_metallib_end:" >> $(TEMP_ASSEMBLY) @$(AS) $(TEMP_ASSEMBLY) -o $@ @rm -f ${TEMP_ASSEMBLY} endif diff --git a/ggml-metal.m b/ggml-metal.m index 3a5476c52..c3451a79b 100644 --- a/ggml-metal.m +++ b/ggml-metal.m @@ -280,6 +280,11 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) { id metal_library; // load library + // + // - first check if the library is embedded + // - then check if the library is in the bundle + // - if not found, load the source and compile it + // - if that fails, return NULL { NSBundle * bundle = nil; #ifdef SWIFT_PACKAGE @@ -287,12 +292,21 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) { #else bundle = [NSBundle bundleForClass:[GGMLMetalClass class]]; #endif + NSError * error = nil; - NSString * libPath = [bundle pathForResource:@"default" ofType:@"metallib"]; - if (libPath != nil) { + +#if GGML_METAL_EMBED_LIBRARY + const bool try_metallib = false; +#else + const bool try_metallib = true; +#endif + + NSString * path_lib = [bundle pathForResource:@"default" ofType:@"metallib"]; + if (try_metallib && path_lib != nil) { // pre-compiled library found - NSURL * libURL = [NSURL fileURLWithPath:libPath]; - GGML_METAL_LOG_INFO("%s: loading '%s'\n", __func__, [libPath UTF8String]); + NSURL * libURL = [NSURL fileURLWithPath:path_lib]; + GGML_METAL_LOG_INFO("%s: loading '%s'\n", __func__, [path_lib UTF8String]); + metal_library = [ctx->device newLibraryWithURL:libURL error:&error]; if (error) { GGML_METAL_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]); @@ -305,31 +319,34 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) { extern const char ggml_metallib_start[]; extern const char ggml_metallib_end[]; - NSString * src = [[NSString alloc] initWithBytes:ggml_metallib_start length:(ggml_metallib_end-ggml_metallib_start) encoding:NSUTF8StringEncoding]; + NSString * src = [[NSString alloc] initWithBytes:ggml_metallib_start length:(ggml_metallib_end-ggml_metallib_start) encoding:NSUTF8StringEncoding]; #else GGML_METAL_LOG_INFO("%s: default.metallib not found, loading from source\n", __func__); - NSString * sourcePath; - NSString * ggmlMetalPathResources = [[NSProcessInfo processInfo].environment objectForKey:@"GGML_METAL_PATH_RESOURCES"]; + NSString * path_source; + NSString * path_resource = [[NSProcessInfo processInfo].environment objectForKey:@"GGML_METAL_PATH_RESOURCES"]; - GGML_METAL_LOG_INFO("%s: GGML_METAL_PATH_RESOURCES = %s\n", __func__, ggmlMetalPathResources ? [ggmlMetalPathResources UTF8String] : "nil"); + GGML_METAL_LOG_INFO("%s: GGML_METAL_PATH_RESOURCES = %s\n", __func__, path_resource ? [path_resource UTF8String] : "nil"); - if (ggmlMetalPathResources) { - sourcePath = [ggmlMetalPathResources stringByAppendingPathComponent:@"ggml-metal.metal"]; + if (path_resource) { + path_source = [path_resource stringByAppendingPathComponent:@"ggml-metal.metal"]; } else { - sourcePath = [bundle pathForResource:@"ggml-metal" ofType:@"metal"]; + path_source = [bundle pathForResource:@"ggml-metal" ofType:@"metal"]; } - if (sourcePath == nil) { + + if (path_source == nil) { GGML_METAL_LOG_WARN("%s: error: could not use bundle path to find ggml-metal.metal, falling back to trying cwd\n", __func__); - sourcePath = @"ggml-metal.metal"; + path_source = @"ggml-metal.metal"; } - GGML_METAL_LOG_INFO("%s: loading '%s'\n", __func__, [sourcePath UTF8String]); - NSString * src = [NSString stringWithContentsOfFile:sourcePath encoding:NSUTF8StringEncoding error:&error]; + + GGML_METAL_LOG_INFO("%s: loading '%s'\n", __func__, [path_source UTF8String]); + + NSString * src = [NSString stringWithContentsOfFile:path_source encoding:NSUTF8StringEncoding error:&error]; if (error) { GGML_METAL_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]); return NULL; } -#endif +#endif // GGML_METAL_EMBED_LIBRARY @autoreleasepool { // dictionary of preprocessor macros diff --git a/ggml-metal.metal b/ggml-metal.metal index ebf2f5b47..63de56325 100644 --- a/ggml-metal.metal +++ b/ggml-metal.metal @@ -4,9 +4,6 @@ #include -#define GGML_COMMON_IMPL_METAL -#include "ggml-common.h" - using namespace metal; #define MAX(x, y) ((x) > (y) ? (x) : (y)) From 68265ebfc6a1bed022973ea0c3145be1450b7e70 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Thu, 14 Mar 2024 12:37:20 +0200 Subject: [PATCH 03/15] embedding : print all resulting embeddings (#899) --- examples/embedding/embedding.cpp | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/examples/embedding/embedding.cpp b/examples/embedding/embedding.cpp index f390c4061..895469a31 100644 --- a/examples/embedding/embedding.cpp +++ b/examples/embedding/embedding.cpp @@ -167,9 +167,9 @@ int main(int argc, char ** argv) { float * out = emb + p * n_embd; batch_decode(ctx, batch, out, s, n_embd); - // print first 3 embeddings + // print the first part of the embeddings fprintf(stdout, "\n"); - for (int j = 0; j < std::min(3, n_prompts); j++) { + for (int j = 0; j < n_prompts; j++) { fprintf(stdout, "embedding %d: ", j); for (int i = 0; i < std::min(16, n_embd); i++) { fprintf(stdout, "%f ", emb[j * n_embd + i]); From 3fe8d7a17f84bd721cd4d8db35365da44b69f68b Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Thu, 14 Mar 2024 12:38:37 +0200 Subject: [PATCH 04/15] ggml : designate enum vals for integer types (#6050) --- ggml.h | 64 +++++++++++++++++++++++++++++----------------------------- 1 file changed, 32 insertions(+), 32 deletions(-) diff --git a/ggml.h b/ggml.h index 1171088a9..ab26c8f59 100644 --- a/ggml.h +++ b/ggml.h @@ -337,24 +337,24 @@ extern "C" { struct ggml_object; struct ggml_context; + // NOTE: always add types at the end of the enum to keep backward compatibility enum ggml_type { - GGML_TYPE_F32 = 0, - GGML_TYPE_F16 = 1, - GGML_TYPE_Q4_0 = 2, - GGML_TYPE_Q4_1 = 3, + GGML_TYPE_F32 = 0, + GGML_TYPE_F16 = 1, + GGML_TYPE_Q4_0 = 2, + GGML_TYPE_Q4_1 = 3, // GGML_TYPE_Q4_2 = 4, support has been removed - // GGML_TYPE_Q4_3 (5) support has been removed - GGML_TYPE_Q5_0 = 6, - GGML_TYPE_Q5_1 = 7, - GGML_TYPE_Q8_0 = 8, - GGML_TYPE_Q8_1 = 9, - // k-quantizations - GGML_TYPE_Q2_K = 10, - GGML_TYPE_Q3_K = 11, - GGML_TYPE_Q4_K = 12, - GGML_TYPE_Q5_K = 13, - GGML_TYPE_Q6_K = 14, - GGML_TYPE_Q8_K = 15, + // GGML_TYPE_Q4_3 = 5, support has been removed + GGML_TYPE_Q5_0 = 6, + GGML_TYPE_Q5_1 = 7, + GGML_TYPE_Q8_0 = 8, + GGML_TYPE_Q8_1 = 9, + GGML_TYPE_Q2_K = 10, + GGML_TYPE_Q3_K = 11, + GGML_TYPE_Q4_K = 12, + GGML_TYPE_Q5_K = 13, + GGML_TYPE_Q6_K = 14, + GGML_TYPE_Q8_K = 15, GGML_TYPE_IQ2_XXS = 16, GGML_TYPE_IQ2_XS = 17, GGML_TYPE_IQ3_XXS = 18, @@ -363,9 +363,9 @@ extern "C" { GGML_TYPE_IQ3_S = 21, GGML_TYPE_IQ2_S = 22, GGML_TYPE_IQ4_XS = 23, - GGML_TYPE_I8, - GGML_TYPE_I16, - GGML_TYPE_I32, + GGML_TYPE_I8 = 24, + GGML_TYPE_I16 = 25, + GGML_TYPE_I32 = 26, GGML_TYPE_COUNT, }; @@ -383,20 +383,20 @@ extern "C" { // model file types enum ggml_ftype { - GGML_FTYPE_UNKNOWN = -1, - GGML_FTYPE_ALL_F32 = 0, - GGML_FTYPE_MOSTLY_F16 = 1, // except 1d tensors - GGML_FTYPE_MOSTLY_Q4_0 = 2, // except 1d tensors - GGML_FTYPE_MOSTLY_Q4_1 = 3, // except 1d tensors + GGML_FTYPE_UNKNOWN = -1, + GGML_FTYPE_ALL_F32 = 0, + GGML_FTYPE_MOSTLY_F16 = 1, // except 1d tensors + GGML_FTYPE_MOSTLY_Q4_0 = 2, // except 1d tensors + GGML_FTYPE_MOSTLY_Q4_1 = 3, // except 1d tensors GGML_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4, // tok_embeddings.weight and output.weight are F16 - GGML_FTYPE_MOSTLY_Q8_0 = 7, // except 1d tensors - GGML_FTYPE_MOSTLY_Q5_0 = 8, // except 1d tensors - GGML_FTYPE_MOSTLY_Q5_1 = 9, // except 1d tensors - GGML_FTYPE_MOSTLY_Q2_K = 10, // except 1d tensors - GGML_FTYPE_MOSTLY_Q3_K = 11, // except 1d tensors - GGML_FTYPE_MOSTLY_Q4_K = 12, // except 1d tensors - GGML_FTYPE_MOSTLY_Q5_K = 13, // except 1d tensors - GGML_FTYPE_MOSTLY_Q6_K = 14, // except 1d tensors + GGML_FTYPE_MOSTLY_Q8_0 = 7, // except 1d tensors + GGML_FTYPE_MOSTLY_Q5_0 = 8, // except 1d tensors + GGML_FTYPE_MOSTLY_Q5_1 = 9, // except 1d tensors + GGML_FTYPE_MOSTLY_Q2_K = 10, // except 1d tensors + GGML_FTYPE_MOSTLY_Q3_K = 11, // except 1d tensors + GGML_FTYPE_MOSTLY_Q4_K = 12, // except 1d tensors + GGML_FTYPE_MOSTLY_Q5_K = 13, // except 1d tensors + GGML_FTYPE_MOSTLY_Q6_K = 14, // except 1d tensors GGML_FTYPE_MOSTLY_IQ2_XXS = 15, // except 1d tensors GGML_FTYPE_MOSTLY_IQ2_XS = 16, // except 1d tensors GGML_FTYPE_MOSTLY_IQ3_XXS = 17, // except 1d tensors From 3ca23481dd309bd51cc31c73a4cc34f922cc372f Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Ond=C5=99ej=20=C4=8Cert=C3=ADk?= Date: Thu, 14 Mar 2024 04:40:14 -0600 Subject: [PATCH 05/15] gguf-py : add support for I8, I16 and I32 (#6045) * Refactor dtype handling to be extensible This code is equivalent as before, but now it is prepared to easily add more NumPy dtypes. * Add support for I8, I16 and I32 These types are allowed in the GGUF specification. * Add support for I8, I16 and I32 to gguf_writer * Add support for I8, I16, I32 to gguf_reader --- gguf-py/gguf/constants.py | 6 ++++++ gguf-py/gguf/gguf_reader.py | 9 +++++++++ gguf-py/gguf/gguf_writer.py | 16 ++++++++++++---- 3 files changed, 27 insertions(+), 4 deletions(-) diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index b23badb10..99f71f0a1 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -661,6 +661,9 @@ class GGMLQuantizationType(IntEnum): IQ3_S = 21 IQ2_S = 22 IQ4_XS = 23 + I8 = 24 + I16 = 25 + I32 = 26 class GGUFEndian(IntEnum): @@ -727,6 +730,9 @@ GGML_QUANT_SIZES = { GGMLQuantizationType.IQ3_S: (256, 2 + QK_K // 4 + QK_K // 8 + QK_K // 32 + 4), GGMLQuantizationType.IQ2_S: (256, 2 + QK_K // 4 + QK_K // 16), GGMLQuantizationType.IQ4_XS: (256, 2 + 2 + QK_K // 2 + QK_K // 64), + GGMLQuantizationType.I8: (1, 1), + GGMLQuantizationType.I16: (1, 2), + GGMLQuantizationType.I32: (1, 4), } diff --git a/gguf-py/gguf/gguf_reader.py b/gguf-py/gguf/gguf_reader.py index 5b6d4ba6b..1c10f5753 100644 --- a/gguf-py/gguf/gguf_reader.py +++ b/gguf-py/gguf/gguf_reader.py @@ -248,6 +248,15 @@ class GGUFReader: elif ggml_type == GGMLQuantizationType.F16: item_count = n_elems item_type = np.float16 + elif ggml_type == GGMLQuantizationType.I8: + item_count = n_elems + item_type = np.int8 + elif ggml_type == GGMLQuantizationType.I16: + item_count = n_elems + item_type = np.int16 + elif ggml_type == GGMLQuantizationType.I32: + item_count = n_elems + item_type = np.int32 else: item_count = n_bytes item_type = np.uint8 diff --git a/gguf-py/gguf/gguf_writer.py b/gguf-py/gguf/gguf_writer.py index e49c5db68..9c1eeac31 100644 --- a/gguf-py/gguf/gguf_writer.py +++ b/gguf-py/gguf/gguf_writer.py @@ -196,9 +196,6 @@ class GGUFWriter: if self.state is not WriterState.EMPTY: raise ValueError(f'Expected output file to be empty, got {self.state}') - if raw_dtype is None and tensor_dtype not in (np.float32, np.float16): - raise ValueError("Only F32 and F16 tensors are supported for now") - encoded_name = name.encode("utf8") self.ti_data += self._pack("Q", len(encoded_name)) self.ti_data += encoded_name @@ -207,7 +204,18 @@ class GGUFWriter: for i in range(n_dims): self.ti_data += self._pack("Q", tensor_shape[n_dims - 1 - i]) if raw_dtype is None: - dtype = GGMLQuantizationType.F32 if tensor_dtype == np.float32 else GGMLQuantizationType.F16 + if tensor_shape == np.float32: + dtype = GGMLQuantizationType.F32 + elif tensor_dtype == np.float16: + dtype = GGMLQuantizationType.F16 + elif tensor_dtype == np.int8: + dtype = GGMLQuantizationType.I8 + elif tensor_dtype == np.int16: + dtype = GGMLQuantizationType.I16 + elif tensor_dtype == np.int32: + dtype = GGMLQuantizationType.I32 + else: + raise ValueError("Only F32, F16, I8, I16, I32 tensors are supported for now") else: dtype = raw_dtype self.ti_data += self._pack("I", dtype) From 2c4fb69246834503db7b78bcbedcef506bbc60c4 Mon Sep 17 00:00:00 2001 From: Michael Podvitskiy Date: Thu, 14 Mar 2024 11:56:48 +0100 Subject: [PATCH 06/15] llama : optimize defrag moves + fix fragmentation calculation (#6037) * attempt to reduce the impact of a worst-case scenario * fragmentation calculation fix * Update llama.cpp --------- Co-authored-by: Georgi Gerganov --- llama.cpp | 30 +++++++++++++++++++----------- 1 file changed, 19 insertions(+), 11 deletions(-) diff --git a/llama.cpp b/llama.cpp index 38e7036a7..ff467c575 100644 --- a/llama.cpp +++ b/llama.cpp @@ -9036,8 +9036,8 @@ static int llama_decode_internal( //llama_synchronize(&lctx); // decide if we need to defrag the kv cache - if (cparams.defrag_thold >= 0.0f) { - const float fragmentation = kv_self.n >= 128 ? 1.0f - float(kv_self.used + n_tokens_all)/float(kv_self.n) : 0.0f; + if (cparams.causal_attn && cparams.defrag_thold >= 0.0f) { + const float fragmentation = kv_self.n >= 128 ? 1.0f - float(kv_self.used)/float(kv_self.n) : 0.0f; // queue defragmentation for next llama_kv_cache_update if (fragmentation > cparams.defrag_thold) { @@ -9069,6 +9069,11 @@ static void llama_kv_cache_defrag_internal(struct llama_context & lctx) { // number of cells moved uint32_t n_moves = 0; + // each move requires 6*n_layer tensors (see build_defrag) + // - source view, destination view, copy operation + // - x2 for keys and values + const uint32_t max_moves = LLAMA_MAX_NODES/(6*n_layer); + // determine which KV cells to move where // // cell i moves to ids[i] @@ -9095,15 +9100,6 @@ static void llama_kv_cache_defrag_internal(struct llama_context & lctx) { nh++; } - // each move requires 6*n_layer tensors (see build_defrag) - // - source view, destination view, copy operation - // - x2 for keys and values - // - if (6*(n_moves + nh)*n_layer >= LLAMA_MAX_NODES) { - // the graph is too big, we cannot move more cells - break; - } - uint32_t nf = 0; uint32_t is = n_kv - 1; @@ -9133,11 +9129,19 @@ static void llama_kv_cache_defrag_internal(struct llama_context & lctx) { // are we moving a continuous block of memory? bool cont = false; + // should we stop searching for the next move? + bool stop = false; + // go back and move the nf cells to the hole for (; i1 < n_kv; ++i1) { auto & cell1 = kv_self.cells[i1]; if (cell1.is_empty() || ids[i1] != n_kv) { + if (n_moves == max_moves) { + stop = true; + break; + } + cont = false; continue; } @@ -9164,6 +9168,10 @@ static void llama_kv_cache_defrag_internal(struct llama_context & lctx) { } } + if (stop || n_moves == max_moves) { + break; + } + //LLAMA_LOG_INFO("(tmp log) KV defrag: move [%u, %u) to [%u, %u)\n", is, i1 + 1, i0, i0 + nh); i0 += nh - 1; From a44bc969e4cd62ca9f4332e17fe3c51f2093e7c6 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Thu, 14 Mar 2024 13:13:06 +0200 Subject: [PATCH 07/15] llama : fix typo --- llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/llama.cpp b/llama.cpp index ff467c575..eb48b1e90 100644 --- a/llama.cpp +++ b/llama.cpp @@ -3932,7 +3932,7 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) { LLAMA_LOG_INFO("%s: n_ff = %u\n", __func__, hparams.n_ff); LLAMA_LOG_INFO("%s: n_expert = %u\n", __func__, hparams.n_expert); LLAMA_LOG_INFO("%s: n_expert_used = %u\n", __func__, hparams.n_expert_used); - LLAMA_LOG_INFO("%s: causal attm = %d\n", __func__, hparams.causal_attn); + LLAMA_LOG_INFO("%s: causal attn = %d\n", __func__, hparams.causal_attn); LLAMA_LOG_INFO("%s: pooling type = %d\n", __func__, hparams.pooling_type); LLAMA_LOG_INFO("%s: rope type = %d\n", __func__, hparams.rope_type); LLAMA_LOG_INFO("%s: rope scaling = %s\n", __func__, rope_scaling_type); From 43241adf22e8231ffaf3827d2c9310cc0ffd5ac5 Mon Sep 17 00:00:00 2001 From: Pierrick Hymbert Date: Thu, 14 Mar 2024 12:15:39 +0100 Subject: [PATCH 08/15] server: disable debug release type sanitizer, simplify trigger (#6047) - increase time out for server - do not fail fast --- .github/workflows/server.yml | 17 ++++++++--------- examples/server/tests/features/steps/steps.py | 9 ++++++++- 2 files changed, 16 insertions(+), 10 deletions(-) diff --git a/.github/workflows/server.yml b/.github/workflows/server.yml index e385e03f3..5e38b3547 100644 --- a/.github/workflows/server.yml +++ b/.github/workflows/server.yml @@ -25,17 +25,14 @@ jobs: strategy: matrix: sanitizer: [ADDRESS, THREAD, UNDEFINED] - build_type: [Debug, Release] + build_type: [Debug] include: - build_type: Release sanitizer: "" - exclude: - - build_type: Release - sanitizer: ADDRESS - - build_type: Release + - build_type: Debug sanitizer: THREAD - - build_type: Release - sanitizer: UNDEFINED + disabled_on_pr: true + fail-fast: false # While -DLLAMA_SANITIZE_THREAD=ON is broken container: image: ubuntu:latest @@ -81,13 +78,14 @@ jobs: - name: Tests id: server_integration_tests + if: ${{ !matrix.disabled_on_pr || !github.event.pull_request }} run: | cd examples/server/tests PORT=8888 ./tests.sh - name: Slow tests id: server_integration_tests_slow - if: ${{ github.event.schedule != '' && matrix.build_type == 'Release' || github.event.inputs.slow_tests == 'true' }} + if: ${{ (github.event.schedule || github.event.inputs.slow_tests == 'true') && matrix.build_type == 'Release' }} run: | cd examples/server/tests PORT=8888 ./tests.sh --stop --no-skipped --no-capture --tags slow @@ -124,13 +122,14 @@ jobs: - name: Tests id: server_integration_tests + if: ${{ !matrix.disabled_on_pr || !github.event.pull_request }} run: | cd examples/server/tests behave.exe --summary --stop --no-capture --exclude 'issues|wrong_usages|passkey' --tags llama.cpp - name: Slow tests id: server_integration_tests_slow - if: ${{ github.event.schedule != '' || github.event.inputs.slow_tests == 'true' }} + if: ${{ (github.event.schedule || github.event.inputs.slow_tests == 'true') && matrix.build_type == 'Release' }} run: | cd examples/server/tests behave.exe --stop --no-skipped --no-capture --tags slow diff --git a/examples/server/tests/features/steps/steps.py b/examples/server/tests/features/steps/steps.py index cfa9f96ec..a59a52d21 100644 --- a/examples/server/tests/features/steps/steps.py +++ b/examples/server/tests/features/steps/steps.py @@ -119,6 +119,10 @@ def step_server_metrics(context): def step_start_server(context): start_server_background(context) attempts = 0 + max_attempts = 20 + if 'GITHUB_ACTIONS' in os.environ: + max_attempts *= 2 + while True: with closing(socket.socket(socket.AF_INET, socket.SOCK_STREAM)) as sock: result = sock.connect_ex((context.server_fqdn, context.server_port)) @@ -126,7 +130,7 @@ def step_start_server(context): print("\x1b[33;46mserver started!\x1b[0m") return attempts += 1 - if attempts > 20: + if attempts > max_attempts: assert False, "server not started" print(f"waiting for server to start, connect error code = {result}...") time.sleep(0.1) @@ -943,6 +947,9 @@ async def wait_for_health_status(context, print(f"Starting checking for health for expected_health_status={expected_health_status}\n") interval = 0.5 counter = 0 + if 'GITHUB_ACTIONS' in os.environ: + timeout *= 2 + async with aiohttp.ClientSession() as session: while True: async with await session.get(f'{base_url}/health', params=params) as health_response: From 15a333260ab637a040ed0864c206a2ceaf806bb8 Mon Sep 17 00:00:00 2001 From: Jian Liao Date: Thu, 14 Mar 2024 04:18:23 -0700 Subject: [PATCH 09/15] readme : improve readme for Llava-1.6 example (#6044) Co-authored-by: Jian Liao --- examples/llava/README.md | 18 +++++++++++++----- 1 file changed, 13 insertions(+), 5 deletions(-) diff --git a/examples/llava/README.md b/examples/llava/README.md index 35e6d9e5d..67cb0f22b 100644 --- a/examples/llava/README.md +++ b/examples/llava/README.md @@ -63,12 +63,20 @@ Now both the LLaMA part and the image encoder is in the `llava-v1.5-7b` director ```console git clone https://huggingface.co/liuhaotian/llava-v1.6-vicuna-7b ``` -2) Use `llava-surgery-v2.py` which also supports llava-1.5 variants pytorch as well as safetensor models: + +2) Install the required Python packages: + +```sh +pip install -r examples/llava/requirements.txt +``` + +3) Use `llava-surgery-v2.py` which also supports llava-1.5 variants pytorch as well as safetensor models: ```console python examples/llava/llava-surgery-v2.py -C -m ../llava-v1.6-vicuna-7b/ ``` - you will find a llava.projector and a llava.clip file in your model directory -3) Copy the llava.clip file into a subdirectory (like vit), rename it to pytorch_model.bin and add a fitting vit configuration to the directory: + +4) Copy the llava.clip file into a subdirectory (like vit), rename it to pytorch_model.bin and add a fitting vit configuration to the directory: ```console mkdir vit cp ../llava-v1.6-vicuna-7b/llava.clip vit/pytorch_model.bin @@ -76,18 +84,18 @@ cp ../llava-v1.6-vicuna-7b/llava.projector vit/ curl -s -q https://huggingface.co/cmp-nct/llava-1.6-gguf/raw/main/config_vit.json -o vit/config.json ``` -4) Create the visual gguf model: +5) Create the visual gguf model: ```console python ./examples/llava/convert-image-encoder-to-gguf.py -m vit --llava-projector vit/llava.projector --output-dir vit --clip-model-is-vision ``` - This is similar to llava-1.5, the difference is that we tell the encoder that we are working with the pure vision model part of CLIP -5) Then convert the model to gguf format: +6) Then convert the model to gguf format: ```console python ./convert.py ../llava-v1.6-vicuna-7b/ --skip-unknown ``` -6) And finally we can run the llava-cli using the 1.6 model version: +7) And finally we can run the llava-cli using the 1.6 model version: ```console ./llava-cli -m ../llava-v1.6-vicuna-7b/ggml-model-f16.gguf --mmproj vit/mmproj-model-f16.gguf --image some-image.jpg -c 4096 ``` From 77178eedc83d49f31bf757d8e12315d76460be78 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Thu, 14 Mar 2024 13:32:14 +0200 Subject: [PATCH 10/15] gguf-py : fix dtype check (#6045) --- gguf-py/gguf/gguf_writer.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/gguf-py/gguf/gguf_writer.py b/gguf-py/gguf/gguf_writer.py index 9c1eeac31..4d389be95 100644 --- a/gguf-py/gguf/gguf_writer.py +++ b/gguf-py/gguf/gguf_writer.py @@ -204,7 +204,7 @@ class GGUFWriter: for i in range(n_dims): self.ti_data += self._pack("Q", tensor_shape[n_dims - 1 - i]) if raw_dtype is None: - if tensor_shape == np.float32: + if tensor_dtype == np.float32: dtype = GGMLQuantizationType.F32 elif tensor_dtype == np.float16: dtype = GGMLQuantizationType.F16 From 044ec4b2a567f649459ccd20af2f387c784faa51 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Thu, 14 Mar 2024 15:14:14 +0200 Subject: [PATCH 11/15] embedding : add EOS token if not present (#899) --- examples/embedding/embedding.cpp | 11 +++++++++-- 1 file changed, 9 insertions(+), 2 deletions(-) diff --git a/examples/embedding/embedding.cpp b/examples/embedding/embedding.cpp index 895469a31..cbf9aa2b5 100644 --- a/examples/embedding/embedding.cpp +++ b/examples/embedding/embedding.cpp @@ -112,13 +112,20 @@ int main(int argc, char ** argv) { // tokenize the prompts and trim std::vector> inputs; for (const auto & prompt : prompts) { - auto inp = ::llama_tokenize(ctx, prompt, true); + auto inp = ::llama_tokenize(ctx, prompt, true, false); if (inp.size() > n_batch) { inp.resize(n_batch); } inputs.push_back(inp); } + // add eos if not present + for (auto & inp : inputs) { + if (inp.empty() || inp.back() != llama_token_eos(model)) { + inp.push_back(llama_token_eos(model)); + } + } + // tokenization stats if (params.verbose_prompt) { for (int i = 0; i < (int) inputs.size(); i++) { @@ -172,7 +179,7 @@ int main(int argc, char ** argv) { for (int j = 0; j < n_prompts; j++) { fprintf(stdout, "embedding %d: ", j); for (int i = 0; i < std::min(16, n_embd); i++) { - fprintf(stdout, "%f ", emb[j * n_embd + i]); + fprintf(stdout, "%9.6f ", emb[j * n_embd + i]); } fprintf(stdout, "\n"); } From 69ff61397d2b7b550dcdda4a35b35128892408b0 Mon Sep 17 00:00:00 2001 From: Michael Podvitskiy Date: Thu, 14 Mar 2024 17:21:56 +0100 Subject: [PATCH 12/15] llama : support models without vocabulary (#5798) * additional methods to read model and ctx parameters * vocab size as a part of a model metadata * models without vocabulary, convert.py part * models without vocabulary, llama.cpp part * PR clean up * converter scrypt fixes * llama_vocab_type update (renamed the new key) * pr review fixes * revert function renaming * one more NoVocab assert --- convert.py | 126 +++++++++++++++++++++--------------- gguf-py/gguf/constants.py | 2 + gguf-py/gguf/gguf_writer.py | 3 + llama.cpp | 92 +++++++++++++++++--------- llama.h | 7 +- 5 files changed, 142 insertions(+), 88 deletions(-) diff --git a/convert.py b/convert.py index c15f8c47e..161430f3e 100755 --- a/convert.py +++ b/convert.py @@ -332,6 +332,9 @@ class Params: # class BpeVocab: + tokenizer_model = "gpt2" + name = "bpe" + def __init__(self, fname_tokenizer: Path, fname_added_tokens: Path | None) -> None: self.bpe_tokenizer = json.loads(open(str(fname_tokenizer), encoding="utf-8").read()) if isinstance(self.bpe_tokenizer.get('model'), dict): @@ -390,6 +393,9 @@ class BpeVocab: class SentencePieceVocab: + tokenizer_model = "llama" + name = "spm" + def __init__(self, fname_tokenizer: Path, fname_added_tokens: Path | None) -> None: self.sentencepiece_tokenizer = SentencePieceProcessor(str(fname_tokenizer)) added_tokens: dict[str, int] @@ -453,6 +459,9 @@ class SentencePieceVocab: class HfVocab: + tokenizer_model = "llama" + name = "hfft" + def __init__(self, fname_tokenizer: Path, fname_added_tokens: Path | None = None) -> None: try: from transformers import AutoTokenizer @@ -553,7 +562,15 @@ class HfVocab: return f"" -Vocab: TypeAlias = "BpeVocab | SentencePieceVocab | HfVocab" +class NoVocab: + tokenizer_model = "no_vocab" + name = "no_vocab" + + def __repr__(self) -> str: + return "" + + +Vocab: TypeAlias = "BpeVocab | SentencePieceVocab | HfVocab | NoVocab" # @@ -935,8 +952,10 @@ def check_vocab_size(params: Params, vocab: Vocab, pad_vocab: bool = False) -> N # Handle special case where the model's vocab size is not set if params.n_vocab == -1: raise ValueError( - f"The model's vocab size is set to -1 in params.json. Please update it manually. Maybe {vocab.vocab_size}?" + f"The model's vocab size is set to -1 in params.json. Please update it manually.{f' Maybe {vocab.vocab_size}?' if hasattr(vocab, 'vocab_size') else ''}" ) + if isinstance(vocab, NoVocab): + return # model has no vocab # Check for a vocab size mismatch if params.n_vocab == vocab.vocab_size: @@ -977,6 +996,7 @@ class OutputFile: name = str(params.path_model.parent).split('/')[-1] self.gguf.add_name (name) + self.gguf.add_vocab_size (params.n_vocab) self.gguf.add_context_length (params.n_ctx) self.gguf.add_embedding_length (params.n_embd) self.gguf.add_block_count (params.n_layer) @@ -1013,21 +1033,9 @@ class OutputFile: if params.ftype is not None: self.gguf.add_file_type(params.ftype) - def handle_tokenizer_model(self, vocab: Vocab) -> str: - # Map the vocab types to the supported tokenizer models - tokenizer_model = { - SentencePieceVocab: "llama", - HfVocab: "llama", - BpeVocab: "gpt2", - }.get(type(vocab)) - - # Block if vocab type is not predefined - if tokenizer_model is None: - raise ValueError("Unknown vocab type: Not supported") - - return tokenizer_model - def extract_vocabulary_from_model(self, vocab: Vocab) -> tuple[list[bytes], list[float], list[gguf.TokenType]]: + assert not isinstance(vocab, NoVocab) + tokens = [] scores = [] toktypes = [] @@ -1043,11 +1051,8 @@ class OutputFile: return tokens, scores, toktypes def add_meta_vocab(self, vocab: Vocab) -> None: - # Handle the tokenizer model - tokenizer_model = self.handle_tokenizer_model(vocab) - # Ensure that tokenizer_model is added to the GGUF model - self.gguf.add_tokenizer_model(tokenizer_model) + self.gguf.add_tokenizer_model(vocab.tokenizer_model) # Extract model vocabulary for model conversion tokens, scores, toktypes = self.extract_vocabulary_from_model(vocab) @@ -1074,6 +1079,26 @@ class OutputFile: def write_tensor_info(self) -> None: self.gguf.write_ti_data_to_file() + def write_tensor_data(self, ftype: GGMLFileType, model: LazyModel, concurrency: int) -> None: + ndarrays_inner = bounded_parallel_map(OutputFile.do_item, model.items(), concurrency=concurrency) + if ftype == GGMLFileType.MostlyQ8_0: + ndarrays = bounded_parallel_map( + OutputFile.maybe_do_quantize, ndarrays_inner, concurrency=concurrency, max_workers=concurrency, + use_processpool_executor=True, + ) + else: + ndarrays = map(OutputFile.maybe_do_quantize, ndarrays_inner) + + start = time.time() + for i, ((name, lazy_tensor), ndarray) in enumerate(zip(model.items(), ndarrays)): + elapsed = time.time() - start + size = ' x '.join(f"{dim:6d}" for dim in lazy_tensor.shape) + padi = len(str(len(model))) + print( + f"[{i + 1:{padi}d}/{len(model)}] Writing tensor {name:38s} | size {size:16} | type {lazy_tensor.data_type.name:4} | T+{int(elapsed):4}" + ) + self.gguf.write_tensor_data(ndarray) + def close(self) -> None: self.gguf.close() @@ -1082,7 +1107,7 @@ class OutputFile: fname_out: Path, params: Params, vocab: Vocab, svocab: gguf.SpecialVocab, endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE, pad_vocab: bool = False, ) -> None: - check_vocab_size(params, vocab, pad_vocab = pad_vocab) + check_vocab_size(params, vocab, pad_vocab=pad_vocab) of = OutputFile(fname_out, endianess=endianess) @@ -1120,8 +1145,11 @@ class OutputFile: # meta data of.add_meta_arch(params) - of.add_meta_vocab(vocab) - of.add_meta_special_vocab(svocab) + if isinstance(vocab, NoVocab): + of.gguf.add_tokenizer_model(vocab.tokenizer_model) + else: + of.add_meta_vocab(vocab) + of.add_meta_special_vocab(svocab) # tensor info for name, lazy_tensor in model.items(): @@ -1131,24 +1159,7 @@ class OutputFile: of.write_tensor_info() # tensor data - ndarrays_inner = bounded_parallel_map(OutputFile.do_item, model.items(), concurrency = concurrency) - if ftype == GGMLFileType.MostlyQ8_0: - ndarrays = bounded_parallel_map( - OutputFile.maybe_do_quantize, ndarrays_inner, concurrency=concurrency, max_workers=concurrency, - use_processpool_executor=True, - ) - else: - ndarrays = map(OutputFile.maybe_do_quantize, ndarrays_inner) - - start = time.time() - for i, ((name, lazy_tensor), ndarray) in enumerate(zip(model.items(), ndarrays)): - elapsed = time.time() - start - size = ' x '.join(f"{dim:6d}" for dim in lazy_tensor.shape) - padi = len(str(len(model))) - print( - f"[{i+1:{padi}d}/{len(model)}] Writing tensor {name:38s} | size {size:16} | type {lazy_tensor.data_type.name:4} | T+{int(elapsed):4}" - ) - of.gguf.write_tensor_data(ndarray) + of.write_tensor_data(ftype, model, concurrency) of.close() @@ -1309,8 +1320,8 @@ class VocabFactory: return vtype, path raise FileNotFoundError(f"Could not find any of {[self._FILES[vt] for vt in vocab_types]}") - def _create_special_vocab(self, vocab: Vocab, vocabtype: str, model_parent_path: Path) -> gguf.SpecialVocab: - load_merges = vocabtype == "bpe" + def _create_special_vocab(self, vocab: Vocab, model_parent_path: Path) -> gguf.SpecialVocab: + load_merges = vocab.name == "bpe" n_vocab = vocab.vocab_size if hasattr(vocab, "vocab_size") else None return gguf.SpecialVocab( model_parent_path, @@ -1319,30 +1330,34 @@ class VocabFactory: n_vocab=n_vocab, ) - def load_vocab(self, vocab_types: list[str], model_parent_path: Path) -> tuple[Vocab, gguf.SpecialVocab]: + def _create_vocab_by_path(self, vocab_types: list[str]) -> Vocab: vocab_type, path = self._select_file(vocab_types) print(f"Loading vocab file {path!r}, type {vocab_type!r}") added_tokens_path = path.parent / "added_tokens.json" - vocab: Vocab if vocab_type == "bpe": - vocab = BpeVocab( + return BpeVocab( path, added_tokens_path if added_tokens_path.exists() else None ) - elif vocab_type == "spm": - vocab = SentencePieceVocab( + if vocab_type == "spm": + return SentencePieceVocab( path, added_tokens_path if added_tokens_path.exists() else None ) - elif vocab_type == "hfft": - vocab = HfVocab( + if vocab_type == "hfft": + return HfVocab( path.parent, added_tokens_path if added_tokens_path.exists() else None ) + raise ValueError(vocab_type) + + def load_vocab(self, vocab_types: list[str], model_parent_path: Path) -> tuple[Vocab, gguf.SpecialVocab]: + vocab: Vocab + if len(vocab_types) == 1 and "no_vocab" in vocab_types: + vocab = NoVocab() else: - raise ValueError(vocab_type) + vocab = self._create_vocab_by_path(vocab_types) # FIXME: Respect --vocab-dir? special_vocab = self._create_special_vocab( vocab, - vocab_type, model_parent_path, ) return vocab, special_vocab @@ -1380,6 +1395,7 @@ def main(args_in: list[str] | None = None) -> None: parser.add_argument("--dump", action="store_true", help="don't convert, just show what's in the model") parser.add_argument("--dump-single", action="store_true", help="don't convert, just show what's in a single model file") parser.add_argument("--vocab-only", action="store_true", help="extract only the vocab") + parser.add_argument("--no-vocab", action="store_true", help="store model without the vocab") parser.add_argument("--outtype", choices=output_choices, help="output format - note: q8_0 may be very slow (default: f16 or f32 based on input)") parser.add_argument("--vocab-dir", type=Path, help="directory containing tokenizer.model, if separate from model file") parser.add_argument("--vocab-type", help="vocab types to try in order, choose from 'spm', 'bpe', 'hfft' (default: spm,hfft)", default="spm,hfft") @@ -1392,6 +1408,10 @@ def main(args_in: list[str] | None = None) -> None: parser.add_argument("--skip-unknown", action="store_true", help="skip unknown tensor names instead of failing") args = parser.parse_args(args_in) + if args.no_vocab: + if args.vocab_only: + raise ValueError("no need to specify --vocab-only if using --no-vocab") + args.vocab_type = "no_vocab" if args.dump_single: model_plus = lazy_load_file(args.model) @@ -1442,7 +1462,7 @@ def main(args_in: list[str] | None = None) -> None: print(f"Wrote {outfile}") return - if model_plus.vocab is not None and args.vocab_dir is None: + if model_plus.vocab is not None and args.vocab_dir is None and not args.no_vocab: vocab = model_plus.vocab print(f"Vocab info: {vocab}") diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index 99f71f0a1..2d7cf16c1 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -32,6 +32,7 @@ class Keys: FILE_TYPE = "general.file_type" class LLM: + VOCAB_SIZE = "{arch}.vocab_size" CONTEXT_LENGTH = "{arch}.context_length" EMBEDDING_LENGTH = "{arch}.embedding_length" BLOCK_COUNT = "{arch}.block_count" @@ -752,6 +753,7 @@ KEY_GENERAL_SOURCE_HF_REPO = Keys.General.SOURCE_HF_REPO KEY_GENERAL_FILE_TYPE = Keys.General.FILE_TYPE # LLM +KEY_VOCAB_SIZE = Keys.LLM.VOCAB_SIZE KEY_CONTEXT_LENGTH = Keys.LLM.CONTEXT_LENGTH KEY_EMBEDDING_LENGTH = Keys.LLM.EMBEDDING_LENGTH KEY_BLOCK_COUNT = Keys.LLM.BLOCK_COUNT diff --git a/gguf-py/gguf/gguf_writer.py b/gguf-py/gguf/gguf_writer.py index 4d389be95..81b2eb884 100644 --- a/gguf-py/gguf/gguf_writer.py +++ b/gguf-py/gguf/gguf_writer.py @@ -321,6 +321,9 @@ class GGUFWriter: self.data_alignment = alignment self.add_uint32(Keys.General.ALIGNMENT, alignment) + def add_vocab_size(self, size: int) -> None: + self.add_uint32(Keys.LLM.VOCAB_SIZE.format(arch=self.arch), size) + def add_context_length(self, length: int) -> None: self.add_uint32(Keys.LLM.CONTEXT_LENGTH.format(arch=self.arch), length) diff --git a/llama.cpp b/llama.cpp index eb48b1e90..10fd53469 100644 --- a/llama.cpp +++ b/llama.cpp @@ -258,6 +258,7 @@ enum llm_kv { LLM_KV_GENERAL_SOURCE_URL, LLM_KV_GENERAL_SOURCE_HF_REPO, + LLM_KV_VOCAB_SIZE, LLM_KV_CONTEXT_LENGTH, LLM_KV_EMBEDDING_LENGTH, LLM_KV_BLOCK_COUNT, @@ -321,6 +322,7 @@ static const std::map LLM_KV_NAMES = { { LLM_KV_GENERAL_SOURCE_URL, "general.source.url" }, { LLM_KV_GENERAL_SOURCE_HF_REPO, "general.source.huggingface.repository" }, + { LLM_KV_VOCAB_SIZE, "%s.vocab_size" }, { LLM_KV_CONTEXT_LENGTH, "%s.context_length" }, { LLM_KV_EMBEDDING_LENGTH, "%s.embedding_length" }, { LLM_KV_BLOCK_COUNT, "%s.block_count" }, @@ -3242,10 +3244,11 @@ static const char * llama_model_type_name(e_model type) { static const char * llama_model_vocab_type_name(enum llama_vocab_type type){ switch (type) { - case LLAMA_VOCAB_TYPE_SPM: return "SPM"; - case LLAMA_VOCAB_TYPE_BPE: return "BPE"; - case LLAMA_VOCAB_TYPE_WPM: return "WPM"; - default: return "unknown"; + case LLAMA_VOCAB_TYPE_NONE: return "no vocab"; + case LLAMA_VOCAB_TYPE_SPM: return "SPM"; + case LLAMA_VOCAB_TYPE_BPE: return "BPE"; + case LLAMA_VOCAB_TYPE_WPM: return "WPM"; + default: return "unknown"; } } @@ -3277,14 +3280,14 @@ static void llm_load_hparams( ml.get_key(LLM_KV_GENERAL_NAME, model.name, false); // get hparams kv - ml.get_arr_n(LLM_KV_TOKENIZER_LIST, hparams.n_vocab); - ml.get_key (LLM_KV_CONTEXT_LENGTH, hparams.n_ctx_train); - ml.get_key (LLM_KV_EMBEDDING_LENGTH, hparams.n_embd); - ml.get_key (LLM_KV_FEED_FORWARD_LENGTH, hparams.n_ff); - ml.get_key (LLM_KV_ATTENTION_HEAD_COUNT, hparams.n_head); - ml.get_key (LLM_KV_BLOCK_COUNT, hparams.n_layer); - ml.get_key (LLM_KV_EXPERT_COUNT, hparams.n_expert, false); - ml.get_key (LLM_KV_EXPERT_USED_COUNT, hparams.n_expert_used, false); + ml.get_key(LLM_KV_VOCAB_SIZE, hparams.n_vocab, false) || ml.get_arr_n(LLM_KV_TOKENIZER_LIST, hparams.n_vocab); + ml.get_key(LLM_KV_CONTEXT_LENGTH, hparams.n_ctx_train); + ml.get_key(LLM_KV_EMBEDDING_LENGTH, hparams.n_embd); + ml.get_key(LLM_KV_FEED_FORWARD_LENGTH, hparams.n_ff); + ml.get_key(LLM_KV_ATTENTION_HEAD_COUNT, hparams.n_head); + ml.get_key(LLM_KV_BLOCK_COUNT, hparams.n_layer); + ml.get_key(LLM_KV_EXPERT_COUNT, hparams.n_expert, false); + ml.get_key(LLM_KV_EXPERT_USED_COUNT, hparams.n_expert_used, false); GGML_ASSERT(hparams.n_expert <= LLAMA_MAX_EXPERTS); GGML_ASSERT(hparams.n_expert_used <= hparams.n_expert); @@ -3645,30 +3648,25 @@ static void llm_load_vocab( const auto kv = LLM_KV(model.arch); - const int token_idx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_LIST).c_str()); - if (token_idx == -1) { - throw std::runtime_error("cannot find tokenizer vocab in model file\n"); - } - - const float * scores = nullptr; - const int score_idx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_SCORES).c_str()); - if (score_idx != -1) { - scores = (const float * ) gguf_get_arr_data(ctx, score_idx); - } - - const int * toktypes = nullptr; - const int toktype_idx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_TOKEN_TYPE).c_str()); - if (toktype_idx != -1) { - toktypes = (const int * ) gguf_get_arr_data(ctx, toktype_idx); - } - // determine vocab type { std::string tokenizer_name; ml.get_key(LLM_KV_TOKENIZER_MODEL, tokenizer_name); - if (tokenizer_name == "llama") { + if (tokenizer_name == "no_vocab") { + vocab.type = LLAMA_VOCAB_TYPE_NONE; + + // default special tokens + vocab.special_bos_id = -1; + vocab.special_eos_id = -1; + vocab.special_unk_id = -1; + vocab.special_sep_id = -1; + vocab.special_pad_id = -1; + vocab.linefeed_id = -1; + + return; + } else if (tokenizer_name == "llama") { vocab.type = LLAMA_VOCAB_TYPE_SPM; // default special tokens @@ -3734,6 +3732,23 @@ static void llm_load_vocab( } } + const int token_idx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_LIST).c_str()); + if (token_idx == -1) { + throw std::runtime_error("cannot find tokenizer vocab in model file\n"); + } + + const float * scores = nullptr; + const int score_idx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_SCORES).c_str()); + if (score_idx != -1) { + scores = (const float * ) gguf_get_arr_data(ctx, score_idx); + } + + const int * toktypes = nullptr; + const int toktype_idx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_TOKEN_TYPE).c_str()); + if (toktype_idx != -1) { + toktypes = (const int * ) gguf_get_arr_data(ctx, toktype_idx); + } + const uint32_t n_vocab = gguf_get_arr_n(ctx, token_idx); vocab.id_to_token.resize(n_vocab); @@ -5023,7 +5038,8 @@ static int llama_model_load(const std::string & fname, llama_model & model, llam llm_load_print_meta(ml, model); - if (model.hparams.n_vocab != model.vocab.id_to_token.size()) { + if (model.vocab.type != LLAMA_VOCAB_TYPE_NONE && + model.hparams.n_vocab != model.vocab.id_to_token.size()) { throw std::runtime_error("vocab size mismatch"); } @@ -9361,26 +9377,32 @@ static enum llama_vocab_type llama_vocab_get_type(const llama_vocab & vocab) { } static bool llama_is_normal_token(const llama_vocab & vocab, llama_token id) { + GGML_ASSERT(vocab.type != LLAMA_VOCAB_TYPE_NONE); return vocab.id_to_token[id].type == LLAMA_TOKEN_TYPE_NORMAL; } static bool llama_is_unknown_token(const llama_vocab & vocab, llama_token id) { + GGML_ASSERT(vocab.type != LLAMA_VOCAB_TYPE_NONE); return vocab.id_to_token[id].type == LLAMA_TOKEN_TYPE_UNKNOWN; } static bool llama_is_control_token(const llama_vocab & vocab, llama_token id) { + GGML_ASSERT(vocab.type != LLAMA_VOCAB_TYPE_NONE); return vocab.id_to_token[id].type == LLAMA_TOKEN_TYPE_CONTROL; } static bool llama_is_byte_token(const llama_vocab & vocab, llama_token id) { + GGML_ASSERT(vocab.type != LLAMA_VOCAB_TYPE_NONE); return vocab.id_to_token[id].type == LLAMA_TOKEN_TYPE_BYTE; } static bool llama_is_user_defined_token(const llama_vocab& vocab, llama_token id) { + GGML_ASSERT(vocab.type != LLAMA_VOCAB_TYPE_NONE); return vocab.id_to_token[id].type == LLAMA_TOKEN_TYPE_USER_DEFINED; } static uint8_t llama_token_to_byte(const llama_vocab& vocab, llama_token id) { + GGML_ASSERT(llama_vocab_get_type(vocab) != LLAMA_VOCAB_TYPE_NONE); GGML_ASSERT(llama_is_byte_token(vocab, id)); const auto& token_data = vocab.id_to_token.at(id); switch (llama_vocab_get_type(vocab)) { @@ -9401,6 +9423,7 @@ static uint8_t llama_token_to_byte(const llama_vocab& vocab, llama_token id) { } static llama_token llama_byte_to_token(const llama_vocab & vocab, uint8_t ch) { + GGML_ASSERT(llama_vocab_get_type(vocab) != LLAMA_VOCAB_TYPE_NONE); static const char * hex = "0123456789ABCDEF"; switch (llama_vocab_get_type(vocab)) { case LLAMA_VOCAB_TYPE_SPM: { @@ -10232,6 +10255,8 @@ static std::vector llama_tokenize_internal(const llama_vocab & } } } break; + case LLAMA_VOCAB_TYPE_NONE: + GGML_ASSERT(false); } return output; @@ -13138,7 +13163,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) { } int32_t llama_n_vocab(const struct llama_model * model) { - return model->vocab.id_to_token.size(); + return model->hparams.n_vocab; } int32_t llama_n_ctx_train(const struct llama_model * model) { @@ -13962,14 +13987,17 @@ float * llama_get_embeddings_seq(struct llama_context * ctx, llama_seq_id seq_id } const char * llama_token_get_text(const struct llama_model * model, llama_token token) { + GGML_ASSERT(model->vocab.type != LLAMA_VOCAB_TYPE_NONE); return model->vocab.id_to_token[token].text.c_str(); } float llama_token_get_score(const struct llama_model * model, llama_token token) { + GGML_ASSERT(model->vocab.type != LLAMA_VOCAB_TYPE_NONE); return model->vocab.id_to_token[token].score; } llama_token_type llama_token_get_type(const struct llama_model * model, llama_token token) { + GGML_ASSERT(model->vocab.type != LLAMA_VOCAB_TYPE_NONE); return model->vocab.id_to_token[token].type; } diff --git a/llama.h b/llama.h index 2d16cc9b9..90aa5372e 100644 --- a/llama.h +++ b/llama.h @@ -59,9 +59,10 @@ extern "C" { typedef int32_t llama_seq_id; enum llama_vocab_type { - LLAMA_VOCAB_TYPE_SPM = 0, // SentencePiece - LLAMA_VOCAB_TYPE_BPE = 1, // Byte Pair Encoding - LLAMA_VOCAB_TYPE_WPM = 2, // WordPiece + LLAMA_VOCAB_TYPE_NONE = 0, // For models without vocab + LLAMA_VOCAB_TYPE_SPM = 1, // SentencePiece + LLAMA_VOCAB_TYPE_BPE = 2, // Byte Pair Encoding + LLAMA_VOCAB_TYPE_WPM = 3, // WordPiece }; // note: these values should be synchronized with ggml_rope From 727107707a73b3dc8a497cf9fc9405722c16dd2b Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Ond=C5=99ej=20=C4=8Cert=C3=ADk?= Date: Thu, 14 Mar 2024 11:57:31 -0600 Subject: [PATCH 13/15] gguf-py : bump version to 0.8.0 (#6060) --- gguf-py/pyproject.toml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/gguf-py/pyproject.toml b/gguf-py/pyproject.toml index 9789c2c87..96396e04e 100644 --- a/gguf-py/pyproject.toml +++ b/gguf-py/pyproject.toml @@ -1,6 +1,6 @@ [tool.poetry] name = "gguf" -version = "0.7.0" +version = "0.8.0" description = "Read and write ML models in GGUF for GGML" authors = ["GGML "] packages = [ From 6e0438da3cc95b89cdbf55f45fa4e324d9076792 Mon Sep 17 00:00:00 2001 From: Steve Grubb Date: Thu, 14 Mar 2024 14:29:32 -0400 Subject: [PATCH 14/15] gguf : fix resource leaks (#6061) There several places where a gguf context is allocated. A call to gguf_free is missing in some error paths. Also on linux, llama-bench was missing a fclose. --- examples/gguf/gguf.cpp | 1 + examples/llama-bench/llama-bench.cpp | 1 + examples/llava/clip.cpp | 4 ++++ examples/train-text-from-scratch/train-text-from-scratch.cpp | 1 + 4 files changed, 7 insertions(+) diff --git a/examples/gguf/gguf.cpp b/examples/gguf/gguf.cpp index e67be4fb2..5444503a5 100644 --- a/examples/gguf/gguf.cpp +++ b/examples/gguf/gguf.cpp @@ -211,6 +211,7 @@ static bool gguf_ex_read_1(const std::string & fname) { for (int j = 0; j < ggml_nelements(cur); ++j) { if (data[j] != 100 + i) { fprintf(stderr, "%s: tensor[%d]: data[%d] = %f\n", __func__, i, j, data[j]); + gguf_free(ctx); return false; } } diff --git a/examples/llama-bench/llama-bench.cpp b/examples/llama-bench/llama-bench.cpp index bf94e7e7a..d6e5e0497 100644 --- a/examples/llama-bench/llama-bench.cpp +++ b/examples/llama-bench/llama-bench.cpp @@ -103,6 +103,7 @@ static std::string get_cpu_info() { } } } + fclose(f); } #endif // TODO: other platforms diff --git a/examples/llava/clip.cpp b/examples/llava/clip.cpp index 6653b815d..2035554ea 100644 --- a/examples/llava/clip.cpp +++ b/examples/llava/clip.cpp @@ -995,6 +995,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { if (!new_clip->ctx_data) { fprintf(stderr, "%s: ggml_init() failed\n", __func__); clip_free(new_clip); + gguf_free(ctx); return nullptr; } @@ -1002,6 +1003,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { if (!fin) { printf("cannot open model file for loading tensors\n"); clip_free(new_clip); + gguf_free(ctx); return nullptr; } @@ -1023,6 +1025,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { if (!fin) { printf("%s: failed to seek for tensor %s\n", __func__, name); clip_free(new_clip); + gguf_free(ctx); return nullptr; } int num_bytes = ggml_nbytes(cur); @@ -1908,6 +1911,7 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i break; default: printf("Please use an input file in f32 or f16\n"); + gguf_free(ctx_out); return false; } diff --git a/examples/train-text-from-scratch/train-text-from-scratch.cpp b/examples/train-text-from-scratch/train-text-from-scratch.cpp index 7eafe8515..7d06e401b 100644 --- a/examples/train-text-from-scratch/train-text-from-scratch.cpp +++ b/examples/train-text-from-scratch/train-text-from-scratch.cpp @@ -711,6 +711,7 @@ static bool load_checkpoint_file(const char * filename, struct my_llama_model * load_checkpoint_gguf(fctx, f_ggml_ctx, model, train); + gguf_free(fctx); return true; } From 4755afd1cbd40d93c017e5b98c39796f52345314 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Thu, 14 Mar 2024 22:58:41 +0200 Subject: [PATCH 15/15] llama : fix integer overflow during quantization (#6063) --- llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/llama.cpp b/llama.cpp index 10fd53469..2c3841974 100644 --- a/llama.cpp +++ b/llama.cpp @@ -11977,7 +11977,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n return new_type; } -static int32_t llama_tensor_quantize_internal(enum ggml_type new_type, const float * f32_data, void * new_data, const int chunk_size, int nrows, int n_per_row, const float * imatrix, std::vector & workers, const int nthread) { +static size_t llama_tensor_quantize_internal(enum ggml_type new_type, const float * f32_data, void * new_data, const int chunk_size, int nrows, int n_per_row, const float * imatrix, std::vector & workers, const int nthread) { std::mutex mutex; int counter = 0; size_t new_size = 0;